Artificial Intelligence & Machine Learning Vocabulary: 100 Key English Terms
Machine learning is built and discussed in English: research papers, library documentation, model cards, experiment logs and team chats all share the same vocabulary, and new terms usually appear in English long before they get a Turkish name. This glossary covers the words AI and ML engineers use every day, from data, training and evaluation to large language models, prompts and putting models into production.
100 terms
12 topics
41 core terms
200 example sentences
Read, understand, keep going
No account needed. Read the meaning, see where the term is used, then study the examples. Click any word in the lesson for help. You can also select a word and press Alt + Enter.
Machine Learning Basics
artificial intelligence
noun
Meaning
The field of computer science that builds systems able to do tasks that normally need human intelligence, such as understanding language, recognizing images or making decisions.
Where you use it
Used for the whole field and, informally, for a single system ('the AI said…'). In daily work people almost always say AI; the full form appears in job titles, strategy documents, laws and formal writing.
Examples
The company is hiring three engineers for its new artificial intelligence team.
The report looks at how hospitals use artificial intelligence to plan staff shifts.
Usage note
AI is pronounced as two letters, /ˌeɪ ˈaɪ/. Because it starts with a vowel sound, say 'an AI model', not 'a AI model'.
Common phrases
AI model · AI system · AI research · AI-powered · generative AI
A branch of artificial intelligence in which computers learn patterns from data and use them to make predictions or decisions, instead of following rules written by hand.
Where you use it
Used as a noun and, very often, before another noun: a machine learning model, engineer or pipeline. In speech and writing it is shortened to ML, and 'AI/ML' is common in job ads.
Examples
She moved from backend development to machine learning two years ago.
We use machine learning to flag suspicious transactions before they are approved.
Usage note
'Learning' here is uncountable: say 'machine learning is', never 'a machine learning' or 'machine learnings'.
The result of training: a program with learned parameters that takes an input, such as an image or a sentence, and produces an output, such as a label, a number or a text.
Where you use it
The central word of the field. You train, evaluate, deploy, serve and retrain a model; models can be small or large, open or closed, and they have versions. Informally people also say 'the model thinks…' or 'the model got it wrong'.
Examples
The new model is twice as fast but slightly less accurate on long documents.
Which model version is running in production right now?
Usage note
The US spelling is 'modeling' and the UK spelling is 'modelling'; ML libraries and papers mostly use the US form.
Common phrases
train a model · deploy a model · model version · model performance · pretrained model
A type of machine learning in which the model learns from examples that come with correct answers (labels), so that it can predict the answer for new examples.
Where you use it
Used to describe both a type of problem and a method. Classification and regression are the two main supervised tasks. In practice the biggest cost is getting enough labeled data.
Examples
Fraud detection is usually framed as a supervised learning problem, with past fraud cases as labels.
Without labeled data, we can't use supervised learning here.
Common phrases
supervised learning task · supervised learning problem · supervised model · self-supervised learning
A task in which a model puts each input into one of a fixed set of categories, for example deciding whether an email is spam or which product category a photo shows.
Where you use it
Classification can be binary (two classes), multi-class (one of many) or multi-label (several at once). The model is called a classifier, and its quality is measured with accuracy, precision, recall and F1 score.
Examples
This is a multi-label classification problem: one ticket can be about billing and delivery at the same time.
For simple classification tasks, a small fine-tuned model is cheaper than an LLM.
Common phrases
binary classification · multi-class classification · image classification · text classification · classification model
A type of machine learning that finds structure in data without correct answers (labels), for example grouping similar customers or spotting unusual events.
Where you use it
Typical methods are clustering, dimensionality reduction and anomaly detection. It helps when labels are expensive or impossible to get. A related idea, self-supervised learning, creates its own labels from the data and is how LLMs are pretrained.
Examples
We used unsupervised learning to group support tickets before anyone had labeled them.
Unsupervised learning flagged a group of payments that later turned out to be fraud.
A task in which a model predicts a number on a continuous scale, such as a house price, a delivery time or tomorrow's temperature.
Where you use it
Contrasted with classification, which predicts categories. Linear regression and logistic regression are classic models, although logistic regression is actually used for classification. Regression quality is measured with error metrics such as MAE or RMSE.
Examples
Predicting delivery time is a regression problem, not a classification one.
A simple linear regression was our baseline, and the neural network beat it by only 4%.
Usage note
In software testing, a regression is a bug that brings back a problem that was already fixed. ML teams use both meanings, so listen for the context.
Common phrases
linear regression · logistic regression · regression model · regression task
A type of machine learning in which an agent learns by trial and error, taking actions in an environment and receiving rewards or penalties for the results.
Where you use it
Shortened to RL. Its key words are agent, environment, action, reward and policy. It is used in game-playing systems, robotics and recommendations, and in the RLHF stage that makes LLMs more helpful.
Examples
The warehouse robots were first trained with reinforcement learning in a simulator.
Designing the reward is the hardest part of any reinforcement learning project.
Common phrases
reinforcement learning agent · deep reinforcement learning · reward function · RL environment
An unsupervised task that groups similar items together without predefined labels, for example sorting customers into segments by how they shop.
Where you use it
Common algorithms are k-means, DBSCAN and hierarchical clustering. Each group is a cluster, and items are said to 'cluster together'. It is often run on embeddings to find themes in large collections of texts or images.
Examples
Clustering the complaints revealed three main themes: delivery, refunds and app crashes.
We ran clustering on the embeddings and then named each cluster by hand.
Describes the best result or method known at a given time, usually measured on a well-known benchmark.
Where you use it
Very common in papers ('achieves state-of-the-art results') and in marketing. It is shortened to SOTA, pronounced 'SO-tuh'. Because the field moves fast, a state-of-the-art model may be overtaken within months.
Examples
The paper claims state-of-the-art results on three Turkish benchmarks.
We don't need a state-of-the-art model for this; a small one is fast and good enough.
Usage note
Use hyphens before a noun ('a state-of-the-art model') but usually not after a verb ('this model is state of the art').
Common phrases
state-of-the-art results · state-of-the-art model · achieve state-of-the-art performance · SOTA
A collection of examples, such as images, texts or table rows, gathered and organized for training, testing or analyzing a model.
Where you use it
Countable: one dataset, two datasets. Engineers collect, clean, split, label and version datasets, and well-known public ones are referred to by name. In ML it is written as one word; 'data set' also appears in statistics.
Examples
We split the dataset into training, validation and test sets.
Is this dataset licensed for commercial use?
Usage note
'Data' can be pronounced /ˈdeɪtə/ or /ˈdætə/; both are correct. In tech, 'data' is usually uncountable ('the data is noisy'), so say 'a dataset', not 'a data'.
Common phrases
public dataset · split a dataset · clean a dataset · labeled dataset · dataset version
The examples a model learns from during training. Their quality, size and variety largely decide how well the model works on new data.
Where you use it
Uncountable, like 'data': 'more training data', not 'more training datas'. It is often discussed in terms of quality, coverage, licensing and privacy: what is in the training data and what is missing from it.
Examples
The model struggles with night photos because there were very few in the training data.
We removed all personal information from the training data before the first run.
Common phrases
collect training data · high-quality training data · in the training data · training data pipeline
One measurable piece of information given to a model as input, for example a customer's age, the number of past orders or the brightness of a pixel.
Where you use it
Common in classic machine learning on tables. Engineers select, create, scale and drop features; 'feature engineering' means designing useful features by hand, and 'feature importance' shows which ones affect predictions most.
Examples
Adding the day of the week as a feature improved the demand forecast.
These two features are almost identical, so we can drop one of them.
Usage note
In product and software talk, a feature is a function of an app ('a new search feature'). In ML it means an input variable; the context tells you which.
The correct answer attached to a training example, such as 'spam' for an email or 'cat' for a photo, which a supervised model learns to predict.
Where you use it
Also a verb: to label data means adding these answers, often done by paid annotators or domain experts. People talk about noisy labels, missing labels and label quality. In statistics the same idea is called the target or dependent variable.
Examples
About five percent of the labels were wrong, so we sent them back for review.
Two radiologists labeled each scan independently.
Usage note
US spelling keeps one l ('labeled', 'labeling'); UK spelling doubles it ('labelled', 'labelling').
Common phrases
noisy labels · label data · class label · label quality
The work of adding labels or notes to raw data, such as drawing boxes around objects in images or marking the topic of each sentence, so that it can be used for training or evaluation.
Where you use it
Also used for the labels themselves ('the annotations are in a JSON file'). The people who do it are annotators; teams write annotation guidelines and measure how often annotators agree. Often used as a near synonym of labeling.
Examples
Annotation of 10,000 X-rays took three doctors almost two months.
We rewrote the annotation guidelines because annotators disagreed on a third of the examples.
The steps that clean and prepare raw data before a model uses it, such as removing duplicates, filling in missing values, resizing images or normalizing text.
Where you use it
Engineers write preprocessing steps, scripts and pipelines. The same preprocessing must be applied during training and in production; differences between the two are a common source of bugs. The verb is preprocess.
Examples
The bug was in preprocessing: production images weren't resized the same way.
On this dataset, preprocessing takes longer than training.
Common phrases
data preprocessing · preprocessing pipeline · preprocessing step · text preprocessing
The correct, verified answers that a model's outputs are compared with, usually created by careful human labeling or by real measurements.
Where you use it
Used as a noun and before other nouns: ground truth labels, ground truth data. People also question whether the ground truth itself is reliable, since human labels can be wrong or disagree. Common in evaluation, computer vision and remote sensing.
Examples
For this test we used biopsy results as the ground truth.
The model disagrees with the ground truth on 6% of the images, and some of those labels are wrong.
Common phrases
ground truth labels · ground truth data · compare against the ground truth · establish the ground truth
Creating extra training examples by making small changes to existing ones, such as flipping or cropping images or rephrasing sentences, so that a model sees more variety.
Where you use it
Very common in computer vision and speech, and increasingly used in NLP. It helps reduce overfitting when data is limited. Engineers must check that a change keeps the label correct: a rotated '6' may turn into a '9'.
Examples
Data augmentation with random crops and color changes added three points of accuracy.
We turned off data augmentation for the validation set.
Common phrases
data augmentation techniques · apply data augmentation · image augmentation · augmentation pipeline
A situation in which some classes in a dataset have far more examples than others, for example 990 normal transactions for every 10 fraudulent ones.
Where you use it
Very common in fraud, medical and defect detection. It makes accuracy misleading, so teams look at precision, recall and F1 instead. Typical fixes are resampling, class weights in the loss function and collecting more examples of the rare class.
Examples
Because of the class imbalance, the model learned to predict 'no defect' for almost everything.
We handled the class imbalance by oversampling the rare class.
Common phrases
severe class imbalance · handle class imbalance · imbalanced dataset · class weights
A mistake in which information that won't be available at prediction time, such as test examples or future data, gets into training and makes results look better than they are.
Where you use it
A classic reason why a model looks excellent in testing and then fails in production. Common causes are duplicates across splits, features that contain the answer, and random splits on time-based data. Also called target leakage or test set contamination.
Examples
99% accuracy on day one usually means data leakage, not a great model.
We found data leakage: the 'refund date' feature only exists after a customer complains.
Usage note
In security and privacy, a data leak means that private data was exposed to outsiders. In ML evaluation, data leakage is a mistake in method.
Common phrases
check for data leakage · cause data leakage · target leakage · test set contamination
A model made of many connected units arranged in layers, loosely inspired by the brain. Each connection has a weight that is adjusted during training.
Where you use it
The basic building block of deep learning. People describe a network by its architecture, its number of layers and its number of parameters. In conversation it is often just 'the network' or 'the net'.
Examples
We replaced the decision tree with a small neural network and gained four points of accuracy.
The neural network has twelve layers and about 80 million parameters.
Common phrases
train a neural network · deep neural network · neural network architecture · feedforward neural network
The part of machine learning that uses neural networks with many layers. It powers modern speech recognition, image recognition and large language models.
Where you use it
Used for the field, its methods and its tools: deep learning frameworks, deep learning models, a deep learning course. It is usually contrasted with 'classical' or 'traditional' machine learning, which uses simpler models and hand-made features.
Examples
For tabular data, gradient-boosted trees often still beat deep learning.
Deep learning needs far more data and compute than our old pipeline did.
Common phrases
deep learning model · deep learning framework · deep learning research · apply deep learning
A number inside a model, such as a weight or a bias, that is learned from data during training. Model size is often given as the number of parameters.
Where you use it
Usually plural and paired with a number: a 7-billion-parameter model, often written 7B. Don't confuse parameters, which the model learns, with hyperparameters, which engineers set before training.
Examples
The smaller model has 1.5 billion parameters and fits on a single GPU.
We froze most parameters and trained only the last two layers.
Usage note
The stress is on the second syllable: pa-RAM-e-ter. Turkish speakers often stress the last syllable, as in 'parametre'.
Common phrases
number of parameters · billion-parameter model · freeze parameters · trainable parameters
A neural network architecture, introduced in 2017, that uses attention to process all parts of a sequence in parallel. It is the basis of today's LLMs.
Where you use it
Used as a noun and before other nouns: a transformer model, transformer layers, a transformer-based system. Engineers talk about encoder-only, decoder-only and encoder-decoder transformers. They are used for images and audio too, not only text.
Examples
We replaced the old recurrent model with a small transformer and cut errors by a third.
Most current LLMs are decoder-only transformers.
Usage note
In electrical engineering a transformer is a device that changes voltage; in ML the word always means this architecture.
Common phrases
transformer model · transformer architecture · decoder-only transformer · transformer-based
One level of a neural network: a group of units that takes the output of the previous level, transforms it and passes the result on to the next one.
Where you use it
Networks have an input layer, hidden layers and an output layer. Engineers add, remove, freeze or unfreeze layers, and the 'deep' in deep learning refers to having many of them. Common types include dense, convolutional and attention layers.
Examples
We froze the first ten layers and trained only the top two.
Adding another hidden layer didn't help; the model just overfit faster.
The learned numbers on the connections of a neural network, which decide how strongly each input affects the next layer. Together they hold what the model has learned.
Where you use it
Usually plural. Engineers load, save, freeze, share and download weights, and 'the weights' often means the trained model file itself. An open-weight model is one whose weights are published, even if its training data and code are not.
Examples
Download the weights from the shared drive and load them into the evaluation script.
They published the weights but not the training data.
Usage note
'Open-weight' and 'open-source' are not the same: open source normally also means that the code is shared and the license allows free use and changes.
Common phrases
model weights · load the weights · open weights · freeze the weights · pretrained weights
A simple mathematical function applied to the output of each unit in a neural network. It adds non-linearity, which lets the network learn complex patterns.
Where you use it
Common examples are ReLU, sigmoid, tanh and GELU. Without activation functions, even a deep network would behave like a single linear layer. Engineers choose or swap the activation function when they design an architecture.
Examples
Try ReLU as the activation function in the hidden layers and sigmoid at the output.
Which activation function does this transformer use?
Common phrases
ReLU activation function · choose an activation function · non-linear activation function · output activation
The method that calculates how much each weight in a neural network contributed to the error, by passing the error backward from the output layer toward the input.
Where you use it
Often shortened to backprop. It works together with gradient descent: backpropagation computes the gradients, then the optimizer updates the weights. Frameworks do it automatically, so engineers rarely write it by hand outside courses and interviews.
Examples
During backpropagation, the gradients in the early layers became almost zero.
Our deep learning course spends a whole week on backpropagation.
Common phrases
backpropagation algorithm · run backpropagation · backpropagation through time · compute gradients
A mechanism that lets a model decide which other parts of its input are most relevant to each part, and give those parts more weight when it builds its output.
Where you use it
The core idea of the transformer. Engineers talk about self-attention, attention heads, attention layers and attention maps. Because its cost grows quickly with input length, a lot of research aims to make attention faster.
Examples
The attention maps suggest the model focuses on the drug name when it answers.
Each layer has 16 attention heads.
Usage note
In general English, 'pay attention' means to focus on something. In ML, attention is a precise technical mechanism, not a human-like ability.
To teach a model by showing it many examples and repeatedly adjusting its parameters so that its outputs get closer to the correct answers.
Where you use it
You train a model on data ('train it on medical images'), for a number of epochs or steps, either from scratch or from a checkpoint. The noun is training: a training run, training job, training loop or training cost.
Examples
We trained the model on two million product reviews.
Training from scratch would take weeks, so we started from a public checkpoint.
Usage note
Turkish 'eğitmek' covers both 'educate' and 'train', but for models English uses only 'train'. 'Educate the model' is wrong.
Common phrases
train a model · train on data · train from scratch · training run · training loop
A formula that measures how far a model's predictions are from the correct answers. Training tries to make this value, called the loss, as small as possible.
Where you use it
People choose, define or switch a loss function, then watch 'the loss' go down on training curves. Common examples are cross-entropy for classification and mean squared error for regression.
Examples
We switched to a weighted loss function because the model was ignoring the rare class.
Which loss function did you use for the ranking model?
Usage note
'Loss' rhymes with 'boss' (/lɔːs/); it is not pronounced like 'lose' (/luːz/).
Common phrases
choose a loss function · minimize the loss · training loss · cross-entropy loss
A problem in which a model learns its training data too closely, including noise and accidents, so it scores well on that data but poorly on new data.
Where you use it
A classic warning sign is training loss going down while validation loss goes up. The verb is overfit: a model overfits, or overfits to the training set. Engineers fight it with more data, regularization, dropout and early stopping.
Examples
Validation loss started rising after epoch 8, a clear sign of overfitting.
With only 300 examples, overfitting is almost guaranteed.
Common phrases
reduce overfitting · prevent overfitting · signs of overfitting · overfit to the training data
One complete pass of the training process through the whole training dataset. Models are usually trained for several epochs.
Where you use it
Used with numbers: train for 10 epochs, after epoch 3, loss per epoch. Very large models may see their data only once, so people count steps or tokens instead. In Turkish teams the English word is used almost always.
Examples
Validation accuracy stopped improving after the fifth epoch.
Each epoch takes about forty minutes on four GPUs.
Usage note
American English says /ˈepək/ (EP-uk); British English usually says /ˈiːpɒk/ (EE-pok). Both are understood.
Common phrases
train for ten epochs · per epoch · number of epochs · after each epoch
A hyperparameter that controls how big a step the model takes each time it updates its parameters. Too high makes training unstable; too low makes it very slow.
Where you use it
Usually the first hyperparameter people tune. It is often changed during training by a learning rate schedule, with a short warm-up at the start and a gradual decay later. Written as LR in notes and code, for example 'LR of 3e-4'.
Examples
The loss exploded after a few hundred steps, so we lowered the learning rate.
We increase the learning rate slowly during the first 1,000 steps as a warm-up.
A setting chosen by engineers before training rather than learned from data, such as the learning rate, the batch size or the number of layers.
Where you use it
You tune, search and log hyperparameters. Hyperparameter tuning or hyperparameter search means trying many combinations and keeping the best one by validation score. In code and chat people often write 'hparams'.
Examples
We ran a hyperparameter search over 40 configurations overnight.
Please log all hyperparameters with each run so we can reproduce the results.
The number of training examples a model processes together before it updates its parameters once.
Where you use it
Limited mainly by GPU memory. Bigger batches use the hardware more efficiently but may need a different learning rate. When memory runs out, engineers reduce the batch size or use gradient accumulation.
Examples
We got an out-of-memory error, so I cut the batch size from 64 to 32.
With a larger batch size, each epoch finishes faster.
Common phrases
increase the batch size · effective batch size · mini-batch · batch size of 32
The basic optimization method behind training: it repeatedly changes each parameter a little in the direction that reduces the loss the most.
Where you use it
In practice people use variants: stochastic gradient descent (SGD), which works on small random batches, and optimizers such as Adam that build on it. The gradients themselves are computed by backpropagation.
Examples
Gradient descent was barely moving because the learning rate was far too small.
In the interview they asked me to explain gradient descent on a whiteboard.
A problem in which a model is too simple or trained too little to capture the patterns in the data, so it performs poorly even on the training data.
Where you use it
The opposite of overfitting. The usual sign is high training loss and high validation loss together. Typical fixes are a bigger model, better features, longer training or less regularization. The verb is underfit.
Examples
Both losses are still high, so this looks like underfitting, not overfitting.
A linear model leads to underfitting on data this complex.
Common phrases
signs of underfitting · fix underfitting · underfit the data
Any technique that stops a model from becoming too complex during training, in order to reduce overfitting and help it work better on new data.
Where you use it
Common forms are L1 and L2 regularization (weight decay), dropout and early stopping. Engineers add, increase or reduce regularization depending on whether the model overfits or underfits. The verb is regularize; the UK spelling is regularisation.
Examples
Stronger regularization closed most of the gap between training and validation accuracy.
We use dropout and weight decay for regularization.
A saved copy of a model's weights, and often its optimizer state, at a certain point in training, so that training can be resumed or that version can be used later.
Where you use it
Engineers save, load, resume from and evaluate checkpoints, often every few thousand steps. 'The best checkpoint' is the one with the best validation score. Published models are often called checkpoints too.
Examples
The job crashed overnight, but we resumed from the last checkpoint.
Which checkpoint should we ship, the one from step 40k or step 60k?
Common phrases
save a checkpoint · resume from a checkpoint · best checkpoint · load the checkpoint
To reach a stable state during training, where the loss stops going down in a meaningful way and more training brings little benefit.
Where you use it
Said of models, training runs and losses: 'the model converged after 20 epochs', 'the loss isn't converging'. The noun is convergence. When training never converges, people check the learning rate, the data and the code.
Examples
The model converged after about 12 epochs, so we stopped there.
With the new learning rate, training converges twice as fast.
Common phrases
converge quickly · fail to converge · training converges · convergence speed
A part of the data kept aside and used only at the end, to measure how well a finished model works on examples it has never seen.
Where you use it
Kept separate from the training and validation sets. The golden rule is never to tune anything on the test set; otherwise the final score looks better than the real performance. Also called the held-out set or test split.
Examples
We ran the model on the test set only once, after all tuning was finished.
Some test set examples were also in the training data, so the score isn't reliable.
Common phrases
held-out test set · evaluate on the test set · test set accuracy · test split
The share of a model's predictions that are correct, out of all its predictions. A model with 90% accuracy is right on nine of every ten examples.
Where you use it
The simplest classification metric and often the first one reported. It can mislead when classes are imbalanced: if 99% of emails are not spam, a model that never says 'spam' still reaches 99% accuracy.
Examples
Accuracy went up from 87 to 91 percent after we cleaned the labels.
Don't report only accuracy here; the classes are very imbalanced.
Usage note
In everyday English 'accuracy' and 'precision' are near synonyms, but in ML they are two different metrics.
Common phrases
improve accuracy · test accuracy · high accuracy · top-1 accuracy
A part of the data kept out of training and used during development to compare models, tune settings and decide when to stop training.
Where you use it
It sits between the training set and the test set. Engineers check validation loss after each epoch and choose hyperparameters by validation score. Also called the dev set, especially in NLP.
Examples
We picked the learning rate that gave the best score on the validation set.
The validation set is too small: one wrong prediction changes accuracy by two points.
Common phrases
validation loss · validation accuracy · hold out a validation set · validation split
Of all the cases a model marked as positive, the share that really were positive. High precision means few false alarms.
Where you use it
Usually discussed together with recall, because improving one often lowers the other. Precision matters most when a false positive is costly, for example blocking a real customer's payment as fraud.
Examples
The spam filter has 98% precision, so almost nothing it blocks is real mail.
We raised the threshold to improve precision, but recall dropped.
Usage note
Turkish sources are not consistent: precision is usually 'kesinlik', but some texts call it 'hassasiyet', a word others use for recall. Many teams simply use the English terms to avoid confusion.
Common phrases
high precision · precision and recall · improve precision · precision at k
Of all the cases that really were positive, the share that the model managed to find. High recall means few missed cases.
Where you use it
Recall matters most when missing a case is costly, for example failing to flag a tumor on a scan. In medicine the same measure is called sensitivity. It is reported together with precision or combined with it in the F1 score.
Examples
For the screening model, we care more about recall than precision.
Recall on the rare class is only 40%, so most of those cases are missed.
Usage note
In everyday English, recall means remembering something, or a company calling back faulty products. As a metric, the stress can fall on either syllable: /ˈriːkɔːl/ or /rɪˈkɔːl/.
Common phrases
high recall · recall rate · precision and recall · recall at k
A single number that combines precision and recall by taking their harmonic mean. It is high only when both precision and recall are high.
Where you use it
A standard classification metric, especially with imbalanced classes where accuracy misleads. With many classes it is averaged in different ways: macro F1 treats every class equally, while micro F1 treats every example equally.
Examples
The new model raised the F1 score from 0.71 to 0.78.
Please also report the macro-averaged F1 score; the small classes matter to the client.
Common phrases
report the F1 score · macro F1 · micro F1 · F1 on the test set
A standard set of tasks and data used to measure models in the same way, so that results from different teams can be compared fairly.
Where you use it
Models are evaluated on, scored on and compared across benchmarks, and a new best result tops a benchmark or a leaderboard. It is also a verb: to benchmark a model. Engineers often warn that high benchmark scores don't always mean good real-world performance.
Examples
It scores well on public benchmarks but struggles with our customers' questions.
We built an internal benchmark from 500 real support conversations.
Common phrases
public benchmark · internal benchmark · benchmark score · top a benchmark · benchmark results
A table that shows, for each true class, how many examples the model assigned to each class, so you can see at a glance which classes it mixes up.
Where you use it
For two classes it has four cells: true positives, false positives, true negatives and false negatives. Precision and recall are calculated from these counts. Teams plot it as a heatmap and look at it in almost every model review.
Examples
The confusion matrix shows that the model often mixes up wolves and huskies.
Can you add a confusion matrix for each language to the report?
An evaluation method that splits the data into several parts, trains on all but one and tests on the remaining part, repeating until every part has been used for testing.
Where you use it
The most common form is k-fold cross-validation, for example 5-fold. It gives a more reliable estimate than a single split, especially on small datasets, but needs k times as much training. It is less common with very large models.
Examples
With only 800 patients, we used 10-fold cross-validation instead of a fixed test set.
The scores vary a lot across the cross-validation folds.
Short for large language model: a very large neural network trained on huge amounts of text to predict the next token, which lets it answer questions, write, summarize and code.
Where you use it
Much more common than the full form in everyday talk. Engineers call an LLM through an API, host an open model themselves or build 'LLM apps'. The plural is LLMs, and you say 'an LLM' because the letter L starts with a vowel sound.
Examples
We use an LLM to draft replies, and a team member approves each one before it goes out.
Open-weight LLMs have become good enough for most of our internal tasks.
Common phrases
call an LLM · open-weight LLM · LLM application · LLM provider · fine-tune an LLM
A small unit of text that a language model reads and writes, often a whole word, part of a word or a punctuation mark. Model limits and prices are counted in tokens.
Where you use it
Used constantly in LLM work: input and output tokens, tokens per second, cost per million tokens, maximum tokens. In English text a token is on average a bit shorter than a word; Turkish, with its long words, usually needs more tokens per word.
Examples
The prompt is about 3,000 tokens, so it fits easily in the context window.
With most providers, output tokens cost more than input tokens.
Usage note
In web security, a token is a piece of data that proves who you are (an access token). In LLM work it means a piece of text.
Common phrases
input tokens · output tokens · tokens per second · token limit · cost per token
A confident but false or made-up output from a generative model, such as an invented fact, quote, source or piece of code.
Where you use it
Teams measure, detect and reduce hallucinations, often with retrieval (RAG), citations and automatic checks. The verb is hallucinate: 'the model hallucinated a paper that doesn't exist'. Some researchers prefer the word 'confabulation'.
Examples
This answer is a hallucination: the two court cases it cites never existed.
Grounding answers in our own documents cut hallucinations by more than half.
Common phrases
reduce hallucinations · hallucination rate · detect hallucinations · the model hallucinated
The maximum amount of text, measured in tokens, that a language model can take into account at once, including the prompt, any documents and its own answer.
Where you use it
Described by its size: a 128k context window, a long context window. When a conversation grows longer than the window, the oldest parts must be cut or summarized. Also called context length.
Examples
The whole contract fits in the context window, so we don't need to split it.
Answer quality drops when we fill the last part of the context window.
Common phrases
context window size · long context window · fit in the context window · 128k context window
AI systems that create new content, such as text, images, audio, video or code, instead of only sorting inputs into classes or predicting values.
Where you use it
Used for both the technology and the industry around it, and shortened to GenAI in business talk. Companies discuss generative AI features, use cases and policies, while engineers usually prefer more precise words such as LLM or diffusion model.
Examples
Legal asked us to write a generative AI policy before the tool goes live.
Most of our generative AI features are built on two external LLM providers.
Common phrases
generative AI tools · generative AI policy · generative AI use case · GenAI
The first and largest training stage, in which a model learns general patterns from a huge amount of mostly unlabeled data before it is adapted to specific tasks.
Where you use it
For LLMs, pretraining usually means predicting the next token on trillions of tokens of text, and it is by far the most expensive stage. It is followed by post-training steps such as fine-tuning and RLHF. Also written pre-training; the adjective is pretrained.
Examples
Pretraining took eight weeks on a few thousand GPUs.
We don't do pretraining ourselves; we fine-tune open models.
Common phrases
pretraining data · pretraining run · large-scale pretraining · pretrained model
A setting that controls how random a generative model's output is. Low values give more predictable, focused answers; high values give more varied and creative ones.
Where you use it
Set for each request, usually between 0 and 1 or 2. Teams use a low temperature for extraction, classification and code, and a higher one for brainstorming. Even at 0, outputs are not always exactly identical. Often adjusted together with top-p.
Examples
Set the temperature to 0 for the invoice extractor; we want output that is as consistent as possible.
At a higher temperature the slogans were more creative, but some made no sense.
Common phrases
set the temperature · low temperature · temperature of 0.7 · sampling temperature
Able to work with more than one type of data, such as text, images, audio and video, like a model that can answer questions about a photo.
Where you use it
Used before nouns: multimodal model, multimodal input, multimodal search. Each type of data is called a modality. Many current LLMs are multimodal and accept images and audio as well as text.
Examples
With a multimodal model, technicians can photograph a broken part and ask what it is.
Our search is now multimodal: users can search with text or with an image.
A system built around an LLM that can plan steps and use tools, such as search, code or company software, to complete a task with little human help.
Where you use it
Engineers build agents, give them tools, limit what they may do and log every action. Related phrases are agentic workflows, tool use and multi-agent systems. Because agents act on real systems, permissions and human approval steps are a big topic.
Examples
The AI agent reads the bug report, finds the right file and opens a draft fix for review.
We don't let the agent send emails unless a person approves them first.
Usage note
In Turkish, 'ajan' can also mean a spy, so some teams prefer 'yapay zeka asistanı' or simply use the English word 'agent'.
Common phrases
build an agent · agentic workflow · tool use · multi-agent system
The text or other input given to a generative model to tell it what to do, such as a question, an instruction or a few examples followed by a task.
Where you use it
Engineers write, test, tweak and version prompts, often as templates with variables. It is also a verb: to prompt a model. In Turkish teams 'prompt' is the everyday word; 'istem' appears mostly in formal texts.
Examples
Adding two examples to the prompt fixed most of the formatting errors.
We keep every prompt in version control, just like code.
Usage note
In everyday English 'prompt' can also be an adjective meaning quick ('a prompt reply'), or a message asking the user to do something ('a login prompt').
Common phrases
write a prompt · prompt template · user prompt · prompt the model
Instructions given to a chat model before the conversation starts, which set its role, rules, tone and limits for all the messages that follow.
Where you use it
Written by developers, not end users, and usually hidden from them. Teams keep it in version control and test changes carefully, because a small edit can change behavior everywhere. It is contrasted with the user prompt or user message.
Examples
The system prompt tells the assistant to answer only questions about our products.
After we shortened the system prompt, responses became faster and cheaper.
Common phrases
write a system prompt · update the system prompt · system prompt instructions · leaked system prompt
Short for retrieval-augmented generation: a method in which a system first searches a document collection for relevant passages and then gives them to an LLM to write the answer.
Where you use it
Pronounced like the word 'rag'. Teams build RAG pipelines over company documents so that answers stay up to date and can cite sources without retraining the model. Quality depends heavily on retrieval: chunking, embeddings and ranking.
Examples
With RAG, the chatbot can answer questions about policies we changed last week.
Most wrong answers came from the retrieval step, not the LLM, so we fixed that part of the RAG pipeline first.
Common phrases
RAG pipeline · RAG system · build a RAG app · RAG evaluation
The practice of designing, testing and improving prompts so that a model gives accurate, useful and consistent outputs for a task.
Where you use it
It includes writing clear instructions, adding examples, setting an output format and testing prompts against an evaluation set. Some people find the name too grand, so 'prompt design' and 'context engineering' are also used.
Examples
A week of prompt engineering gave us most of the gains we had expected from fine-tuning.
Good prompt engineering starts with a test set, not with clever wording.
Giving a model a few worked examples of a task inside the prompt, so that it can follow the same pattern for a new input without any extra training.
Where you use it
Contrasted with zero-shot prompting, where you give only instructions and no examples. People say 'add a few shots' or 'a three-shot prompt'. The examples should be varied and match the real input format closely.
Examples
Few-shot prompting with five labeled tickets worked better than a long list of rules.
Zero-shot was fine for English, but Turkish needed few-shot prompting.
Step-by-step reasoning that a model writes out before giving its final answer, which often improves results on math, logic and multi-step problems.
Where you use it
Written with hyphens before a noun: chain-of-thought prompting, chain-of-thought reasoning. Shortened to CoT. It was first triggered with prompts such as 'think step by step'; many newer models are trained to reason this way on their own.
Examples
Chain-of-thought prompting fixed most of the arithmetic mistakes.
We hide the chain of thought from users and show only the final answer.
A database designed to store embeddings and quickly find the ones most similar to a query, which makes it a common part of semantic search and RAG systems.
Where you use it
You index documents, upsert vectors and run similarity or nearest-neighbor searches. Some teams use a dedicated vector database; others add vector search to a database they already have. Also called a vector store.
Examples
All 200,000 help articles are chunked, embedded and stored in the vector database.
Do we really need a separate vector database, or can our existing database handle it?
Common phrases
similarity search · vector store · nearest-neighbor search · index documents
Extra training of an already pretrained model on a smaller, task-specific dataset, so that it does better on that task or follows a particular style.
Where you use it
The noun and adjective take a hyphen (fine-tuning, a fine-tuned model); the verb is fine-tune. Teams compare fine-tuning with prompting and RAG, because it costs more but can give more consistent results.
Examples
Fine-tuning on 5,000 support tickets taught the model our tone of voice.
Before we try fine-tuning, let's see how far we get with better prompts.
Common phrases
fine-tuning data · full fine-tuning · fine-tuning job · fine-tune a model
Reusing what a model learned on one task or dataset as the starting point for a different but related task, instead of training a new model from scratch.
Where you use it
The standard approach when data is limited: take a pretrained model, replace or add a few layers and fine-tune it on your own data. Fine-tuning is the most common form of transfer learning.
Examples
Thanks to transfer learning, 2,000 labeled images were enough to reach 95% accuracy.
We used transfer learning from a general image model to detect defects in steel.
Common phrases
use transfer learning · transfer learning approach · pretrained backbone · domain adaptation
A model that has only been pretrained, before any fine-tuning or instruction training. It continues text well but does not reliably follow instructions or chat.
Where you use it
Contrasted with instruct or chat models, which have been tuned to follow instructions. Teams pick a base model as the starting point for their own fine-tuning. In general talk, 'base model' can also just mean the model you start from.
Examples
We fine-tuned the base model, not the chat version, because we needed a very specific output format.
The base model just continued my question instead of answering it.
Usage note
'Foundation model' is a broader term for any large pretrained model that many applications are built on; it is not the same as 'base model'.
Common phrases
choose a base model · base model checkpoint · base versus instruct model · fine-tune the base model
Short for low-rank adaptation: a fine-tuning method that keeps the original weights frozen and trains only small added matrices, which saves a lot of memory and storage.
Where you use it
The most popular parameter-efficient fine-tuning (PEFT) method. Teams train a LoRA adapter for each task or customer and swap adapters on top of one base model. QLoRA combines it with quantization so that it fits on smaller GPUs.
Examples
With LoRA we fine-tuned the 8B model on a single GPU in one afternoon.
Each customer has its own LoRA adapter, but they all share the same base model.
Usage note
Don't confuse it with LoRa (with a lowercase a), a long-range radio technology used in the Internet of Things.
Common phrases
LoRA adapter · train a LoRA · LoRA rank · QLoRA · merge the LoRA weights
Short for reinforcement learning from human feedback: a training stage in which people rank model answers, a reward model learns their preferences, and the LLM is then tuned to score higher.
Where you use it
Part of post-training, after pretraining and supervised fine-tuning. It played a big role in making chat models more helpful and polite. Related methods include RLAIF, where AI feedback replaces some human feedback, and DPO, which learns from preferences without a separate reward model.
Examples
After RLHF, our testers rated the model's answers as much more helpful.
We worked with medical experts to rank answers for the RLHF data.
Common phrases
RLHF stage · reward model · human preference data · preference tuning
Training a smaller 'student' model to copy the outputs of a larger 'teacher' model, so that it keeps much of the quality while being faster and cheaper to run.
Where you use it
Often just called distillation; the verb is distill, as in 'we distilled the 70B model into a 3B model'. Used to cut inference costs or to run models on phones. Licenses matter here, because some model licenses forbid using outputs to train other models.
Examples
Knowledge distillation gave us a model ten times smaller with only a 2% drop in accuracy.
We used knowledge distillation to build an on-device version of the classifier.
Common phrases
teacher model · student model · distill a model · distillation loss
The area of AI that teaches computers to understand images and video, for example recognizing objects, reading text in photos or tracking movement.
Where you use it
Used for the field and in job titles (computer vision engineer). Typical tasks are image classification, object detection, segmentation and tracking, in areas such as manufacturing, medical imaging, retail and self-driving cars. The short form CV can be confused with CV meaning a résumé.
Examples
Our computer vision system checks every bottle on the line for cracks.
Most of our computer vision models run on the cameras themselves, not in the cloud.
Short for convolutional neural network: a network that slides small learned filters over an image to detect patterns such as edges, textures and shapes.
Where you use it
For years the standard model in computer vision, and still widely used on phones, cameras and embedded devices because it is efficient. It is often compared with vision transformers. Also called a ConvNet.
Examples
A small CNN runs on the doorbell camera and detects people in real time.
We compared a CNN with a vision transformer, and the CNN was faster at similar accuracy.
A computer vision task that finds objects in an image or video and gives both their class, such as 'car' or 'person', and their position, usually as a box.
Where you use it
Used in traffic cameras, retail shelves, factory inspection and self-driving cars. Engineers talk about detectors, bounding boxes, confidence scores and mAP (mean average precision). It goes further than image classification, which gives one label per image.
Examples
Our object detection model misses people who are partly hidden behind cars.
For object detection we need boxes, not just one label per photo.
A computer vision task that labels every pixel of an image with a class or an object, giving exact shapes instead of rough boxes.
Where you use it
Semantic segmentation labels each pixel by class (road, sky, person); instance segmentation also separates individual objects (person 1, person 2). Used in medical imaging, satellite images and self-driving cars. The output is usually called a mask.
Examples
We use image segmentation to measure the exact size of each tumor on the scan.
Labeling data for image segmentation takes much longer than drawing boxes.
A rectangle drawn around an object in an image and stored as coordinates, showing where the object is. It is the standard output of object detection.
Where you use it
Annotators draw bounding boxes, and models predict them together with a class and a confidence score. Box quality is measured with IoU (intersection over union). In code it is often shortened to bbox.
Examples
The bounding boxes are drawn too loosely, so the model learns a lot of background.
Each bounding box is saved as x, y, width and height.
A generative model that learns to remove noise step by step, so it can create new images, audio or video by starting from pure noise and gradually refining it.
Where you use it
The main technology behind most text-to-image generators. Engineers talk about sampling steps, guidance scale, image-to-image editing and inpainting. Generation is slower than with single-pass models because it takes many steps.
Examples
The marketing team uses a diffusion model to show our products in different settings.
Cutting the sampling steps from 50 to 20 made the diffusion model much faster with little loss in quality.
The area of AI that deals with human language, building systems that read, understand, translate, summarize or generate text and speech.
Where you use it
Almost always shortened to NLP in speech and job titles (NLP engineer, NLP team). Classic NLP tasks include sentiment analysis, named entity recognition and machine translation; since LLMs, many of them are handled by one general model.
Examples
Turkish is harder for natural language processing because one word can carry many suffixes.
She wrote her master's thesis on natural language processing for legal documents.
Usage note
Don't confuse it with neuro-linguistic programming, a self-help method that is also shortened to NLP.
A list of numbers (a vector) that represents a word, sentence, image or other item in such a way that items with similar meanings end up close to each other.
Where you use it
Usually plural. You compute, store and compare embeddings, typically with cosine similarity. They power semantic search, recommendations, clustering and RAG systems, and are kept in vector databases.
Examples
We generate an embedding for every help article and store it in the vector database.
Search results improved after we switched to a multilingual embedding model.
Common phrases
generate embeddings · embedding model · text embeddings · embedding vector · embedding space
The component that splits text into tokens and turns them into numbers a model can read, and that turns the model's output numbers back into text.
Where you use it
Every model comes with its own tokenizer, and using the wrong one breaks the results. Engineers look at a tokenizer's vocabulary size and how efficiently it handles a language; some split Turkish words into many small pieces. The process is called tokenization.
Examples
This tokenizer splits long Turkish words into many pieces, so our prompts cost more.
Make sure you load the tokenizer from the same checkpoint as the model.
Usage note
The UK spelling is tokeniser, but code and documentation almost always use the US form tokenizer.
An NLP task that detects the opinion or emotion in a text, for example whether a product review is positive, negative or neutral.
Where you use it
Used on reviews, social media posts, surveys and support chats. It can be simple (positive or negative) or more detailed, by aspect: 'the battery is great but the screen is dim'. Sarcasm and mixed opinions make it hard.
Examples
Sentiment analysis shows that complaints about delivery doubled after the price change.
Our sentiment analysis model struggles with sarcasm in Turkish tweets.
Common phrases
sentiment analysis model · run sentiment analysis · aspect-based sentiment analysis · sentiment score
An NLP task that finds the names of people, organizations and places, as well as dates, amounts and similar items, in a text and labels each one with its type.
Where you use it
Shortened to NER. Used to pull structured data out of contracts, invoices, news and medical notes, and to hide personal data. Custom entity types, such as product codes or drug names, often need extra labeled data.
Examples
Named entity recognition pulls the company names and amounts out of each invoice.
Our named entity recognition model confuses some Turkish surnames with place names.
Common phrases
NER model · entity types · extract entities · named entities
A large, organized collection of texts or speech recordings used to train language models or to study how a language is used.
Where you use it
Engineers build, clean, filter and deduplicate a corpus, and describe it by its size in tokens, its languages and its sources. The word is more common in NLP and linguistics than in other areas of ML.
Examples
We built a Turkish legal corpus of about two billion tokens.
Half of the web corpus was duplicates, so we removed them before training.
Usage note
The plural is irregular: one corpus, two corpora. 'Corpus' is not 'corpse' (a dead body), although both come from the Latin word for 'body'.
Common phrases
text corpus · training corpus · build a corpus · parallel corpus · web corpus
Systematic unfairness in a model's outputs, for example working worse for some groups of people, usually caused by skewed data or design choices.
Where you use it
Teams detect, measure, audit and reduce bias, and describe it in model cards. Common phrases are 'gender bias', 'data bias' and 'bias in hiring models'. In technical papers the word can also have a purely mathematical meaning, so check the context.
Examples
The audit found a clear bias: the model rejected older applicants more often.
We added more speakers with regional accents to reduce bias in speech recognition.
Usage note
'Bias' also has neutral technical meanings in ML: the extra learned number in each neuron (the bias term) and one kind of error in the bias-variance trade-off. Neither is about fairness.
The goal that a model treats different groups of people equitably, for example by giving similar error rates or approval rates across genders, ages or regions.
Where you use it
Teams evaluate fairness with metrics that compare groups and report it in model cards and audits. There are several definitions of fairness and they can't always be met at the same time, so the choice depends on the use case and the law.
Examples
Before launch, we checked fairness across age groups and regions.
The fairness review found that error rates were twice as high for one dialect.
How well people can understand why a model made a particular prediction or decision, often with the help of tools that show which inputs mattered most.
Where you use it
In research it appears as XAI (explainable AI). It matters in credit, insurance, hiring and medicine, where people may need a reason for a decision. Common methods include SHAP and LIME. Often used alongside interpretability, which focuses on how a model works inside.
Examples
For loan decisions explainability matters a lot, because customers can ask for the main reasons behind a rejection.
The explainability report shows that the postcode had too much influence on the score.
Common phrases
model explainability · explainability tools · explainable AI · feature attribution
A rule or check placed around a model to stop harmful, off-topic or unsafe inputs and outputs, for example a filter that blocks personal data in answers.
Where you use it
Usually plural: teams add, test and tune guardrails on both the input and the output side. Guardrails can be simple rules, classifier models or a second LLM that checks the first. The word comes from the metal barriers along roads.
Examples
The guardrails block any answer that contains a credit card number.
After launch we retuned the guardrails, because they were also blocking normal medical questions.
Common phrases
add guardrails · safety guardrails · input and output guardrails · guardrail model
Testing a model or AI system by deliberately trying to make it fail or behave harmfully, in order to find weaknesses before real users or attackers do.
Where you use it
Done by internal teams, outside experts or automated tools, before and after launch. The people involved are red teamers, and their findings feed into guardrails and further training. The term comes from security exercises in which one team plays the attacker.
Examples
Red teaming showed that the assistant would reveal its system prompt if asked in French.
We planned two weeks of red teaming before the public launch.
Common phrases
red teaming exercise · red-team a model · red teamers · automated red teaming
Making an AI system's behavior match human intentions and values, so that it is helpful and honest and avoids causing harm. Also the research field that studies this.
Where you use it
Used both for practical work (RLHF, safety training, evaluations) and for long-term research on more capable systems. People say a model is well aligned or misaligned, and talk about alignment teams and alignment research.
Examples
Alignment work made the model more willing to admit when it doesn't know something.
She left product engineering to join the alignment research team.
Usage note
In general and engineering English, alignment also means putting things in a straight line, or the alignment of text in a document.
Common phrases
alignment research · alignment team · well aligned · misaligned model · safety and alignment
A prompt or technique designed to trick a model into ignoring its safety rules and producing content it is meant to refuse.
Where you use it
Also a verb: to jailbreak a model. Safety teams collect known jailbreaks, test new model versions against them and fix weaknesses. Related to, but different from, prompt injection, where hidden instructions inside data take control of an app.
Examples
The new version resists every jailbreak in our test suite except two.
Within an hour of launch, someone had found a jailbreak for the demo bot.
Usage note
The word comes from phones, where jailbreaking means removing the maker's software limits. It is informal, but it is used in research papers too.
Common phrases
jailbreak attempt · jailbreak a model · known jailbreaks · jailbreak prompt
The stage in which a trained model is used to produce outputs for new inputs, for example answering a user's question or classifying a new photo.
Where you use it
Contrasted with training. Engineers run inference, speed it up and talk about inference cost, inference servers and batch versus real-time inference. It comes up constantly in infrastructure and budget discussions.
Examples
Now that the app has a million users, inference costs us more than training.
We run inference on CPUs at night and on GPUs during peak hours.
Usage note
In statistics and logic, inference means drawing conclusions from evidence. In ML engineering it usually just means running the model.
To put a trained model into a real system so that users or other software can send it requests and receive predictions.
Where you use it
You deploy a model to production, to a server, to the cloud or to a device. The noun is deployment. Teams talk about deploying a new version, rolling it out gradually and rolling it back if something breaks.
Examples
We deployed the new ranking model on Tuesday, and click-through went up by 3%.
Can we deploy the model on the phone itself, or is it too big?
Common phrases
deploy a model · deploy to production · model deployment · on-device deployment
Short for graphics processing unit: a chip that does thousands of simple calculations in parallel, which makes it the main hardware for training and running neural networks.
Where you use it
Engineers talk about GPU memory, GPU hours, GPU clusters and GPU utilization, and about renting GPUs in the cloud. Model sizes are often discussed in terms of how many GPUs they need. Related chips include TPUs and other AI accelerators.
Examples
The 70B model needs at least two of our GPUs just to load.
GPU utilization is only 30%, so the data loader is probably the bottleneck.
The time between sending a request to a model and receiving its response, usually measured in milliseconds or seconds.
Where you use it
Teams track average latency and tail latency such as p95 or p99. For LLMs they also measure time to first token. Latency is often traded off against throughput, cost and model quality.
Examples
The bigger model is more accurate, but latency went from 300 milliseconds to two seconds.
Our p99 latency spikes every time the cache is cleared.
Usage note
Don't confuse latency (how long one request takes) with throughput (how many requests the system can handle per second).
Common phrases
low latency · p99 latency · reduce latency · latency budget · time to first token
A URL where a deployed model can be reached over a network: other software sends a request to it and receives the model's prediction or answer.
Where you use it
You create, call, scale and monitor an endpoint. Cloud platforms offer managed endpoints, and LLM providers expose endpoints such as a chat completions endpoint. Teams often keep separate endpoints for staging and production.
Examples
The fraud model sits behind a REST endpoint that the payment service calls.
The endpoint scaled down to zero overnight, so the first request this morning was slow.
The practices and tools for running machine learning reliably in production: versioning data and models, automating training and deployment, and monitoring models over time.
Where you use it
Built on DevOps ideas. Used for a team (the MLOps team), a role (MLOps engineer) and a set of tools: experiment tracking, model registry, feature store, CI/CD for models and monitoring. For LLM systems some people say LLMOps.
Examples
Without proper MLOps, nobody knew which model version was in production.
Our MLOps pipeline retrains the model every Monday and deploys it if the tests pass.
A change over time in the data a deployed model receives, compared with the data it was trained on, which can slowly make its predictions worse.
Where you use it
Teams monitor for data drift, set alerts and retrain when it grows too large. A related idea is concept drift, where the relationship between the inputs and the correct answer changes, as when customer habits shift after a big event.
Examples
Data drift started when the app launched in a new country with different customers.
The data drift dashboard shows that the average order size has doubled since January.
Common phrases
detect data drift · monitor for data drift · concept drift · drift alert
Storing a model's weights, and sometimes its activations, with fewer bits, for example 8-bit or 4-bit numbers instead of 16-bit, to make it smaller and faster.
Where you use it
Used to run large models on fewer or smaller GPUs, on laptops and on phones. There is usually a small loss in quality, which teams measure before shipping. People say a model is quantized to 4 bits and name formats such as INT8.
Examples
After 4-bit quantization the model fits on a single GPU, and accuracy dropped by less than a point.
Quantization made inference almost twice as fast on our servers.
Usage note
The UK spelling is quantisation. In Turkish teams people often say 'modeli kuantize ettik' informally.
To make a new model, feature or version available to users, often in stages, starting with a small share of traffic and growing as long as the metrics look good.
Where you use it
The noun is written as one word: a rollout. Teams talk about a gradual or staged rollout, rolling out to 5% of users and pausing a rollout. The opposite action is roll back, which means returning to the previous version.
Examples
We'll roll out the new recommendation model to 10% of users first.
The model was rolled out in Turkey last week and in Germany this week.
Usage note
The phrasal verb is two words (roll out) and the noun is one word (rollout). The same pattern applies to set up and setup.
Common phrases
roll out gradually · staged rollout · roll out to all users · roll back
Ten questions from this glossary. Pick the right meaning for each term and see your score.
How to learn artificial intelligence & machine learning vocabulary
Read real texts from the field regularly: paper abstracts, model cards, library documentation, release notes and the issue threads of open-source ML projects.
Learn terms in pairs that are often confused: training and inference, validation set and test set, precision and recall, overfitting and underfitting.
Notice the verbs that go with each noun: you train, evaluate and deploy a model, run inference, tune hyperparameters and label data.
Watch countable and uncountable nouns: 'data' and 'training data' take no 'a', but you can have 'a dataset' or 'two datasets'.
Say abbreviations the way engineers do: LLM, GPU and NLP letter by letter, but RAG and LoRA as single words.
Frequently asked questions
What is the difference between AI, machine learning and deep learning?
Artificial intelligence is the widest term: any system that performs tasks we associate with human intelligence. Machine learning is a part of AI in which systems learn patterns from data instead of following hand-written rules. Deep learning is a part of machine learning that uses neural networks with many layers; today's large language models and image generators are deep learning models.
What is the difference between training and inference?
Training is the phase in which a model learns: it sees data, makes mistakes and adjusts its parameters, often for hours or weeks on many GPUs. Inference is the phase in which the finished model is used: it receives new input and produces a prediction or an answer. Engineers talk about training cost and inference cost separately because they are planned and paid for differently.
Do Turkish AI engineers use the Turkish or the English terms at work?
Usually both. In meetings and chats, English words are often used with Turkish endings: 'modeli fine-tune ettik', 'prompt'u güncelledim', 'inference süresi uzadı'. Turkish equivalents such as 'makine öğrenmesi', 'sinir ağı', 'aşırı öğrenme' and 'doğruluk' are common in university courses, theses, formal reports and articles for the general public. Knowing both helps you move between a Turkish meeting and an English codebase or paper.
How do I pronounce common AI abbreviations like LLM, GPU, RAG and LoRA?
Most are read letter by letter in English: LLM is 'el-el-em', GPU is 'jee-pee-you', NLP is 'en-el-pee' and AI is 'ay-eye'. A few are read as words: RAG sounds like the English word 'rag', and LoRA is usually said 'LOR-uh'. Because AI starts with a vowel sound, English speakers say 'an AI model', not 'a AI model'.
